Original Input Features Research Articles

A PV cell defect detector combined with transformer and attention mechanism

Automated defect detection in electroluminescence (EL) images of photovoltaic (PV) modules on production lines remains a significant challenge, crucial for replacing labor-intensive and costly manual inspections and enhancing production capacity. This paper presents a novel PV defect detection algorithm that leverages the YOLO architecture, integrating an attention mechanism and the Transformer module. We introduce a polarized self-attention mechanism in the feature extraction stage, enabling separate extraction of spatial and semantic features of PV modules, combined with the original input features, to enhance the network’s feature representation capabilities. Subsequently, we integrate the proposed CNN Combined Transformer (CCT) module into the model. The CCT module employs the transformer to extract contextual semantic information more effectively, improving detection accuracy. The experimental results demonstrate that the proposed method achieves a 77.9% mAP50 on the PVEL-AD dataset while preserving real-time detection capabilities. This method enhances the mAP50 by 17.2% compared to the baseline, and the mAP50:95 metric exhibits an 8.4% increase over the baseline.

Research and application of a novel graph convolutional RVFL and evolutionary equilibrium optimizer algorithm considering spatial factors in ultra-short-term solar power prediction

With the increasing proportion of photovoltaic power generation, solar power prediction systems have become more important. Accurate photovoltaic power prediction can assist the grid dispatch department in formulating effective power scheduling plans, improving grid stability and the capacity to accommodate solar energy. To address the challenges in solar power generation forecasting, the paper proposes a new neural network model called Graph Convolutional Random Vector Functional Link (GCRVFL). The output layer of this model includes both nonlinear transformed features from the hidden layer and the original input features, while also considering the spatial characteristics among photovoltaic sites. Furthermore, Singular Spectrum Analysis (SSA) is utilized to decompose the original solar power time series, extracting different component sequences from the photovoltaic power time series. Opposite based learning and cross-mutation are employed to improve the original Equilibrium Optimizer (EO) algorithm, resulting in an Evolutionary Equilibrium Optimizer (EEO) algorithm. The EEO algorithm is employed to optimize the weight threshold of GCRVFL, leading to a novel photovoltaic power prediction model called SSA-EEO-GCRVFL. Through four datasets and eight control experiments, it can be observed that the proposed model achieves the best predictive performance and is capable of fulfilling the task of photovoltaic power prediction.

Detection Based on Semantics and a Detail Infusion Feature Pyramid Network and a Coordinate Adaptive Spatial Feature Fusion Mechanism Remote Sensing Small Object Detector

In response to the challenges of remote sensing imagery, such as unmanned aerial vehicle (UAV) aerial imagery, including differences in target dimensions, the dominance of small targets, and dense clutter and occlusion in complex environments, this paper optimizes the YOLOv8n model and proposes an innovative small-object-detection model called DDSC-YOLO. First, a DualC2f structure is introduced to improve the feature-extraction capabilities of the model. This structure uses dual-convolutions and group convolution techniques to effectively address the issues of cross-channel communication and preserving information in the original input feature mappings. Next, a new attention mechanism, DCNv3LKA, was developed. This mechanism uses adaptive and fine-grained information-extraction methods to simulate receptive fields similar to self-attention, allowing adaptation to a wide range of target size variations. To address the problem of false and missed detection of small targets in aerial photography, we designed a Semantics and Detail Infusion Feature Pyramid Network (SDI-FPN) and added a dedicated detection scale specifically for small targets, effectively mitigating the loss of contextual information in the model. In addition, the coordinate adaptive spatial feature fusion (CASFF) mechanism is used to optimize the original detection head, effectively overcoming multi-scale information conflicts while significantly improving small target localization accuracy and long-range dependency perception. Testing on the VisDrone2019 dataset shows that the DDSC-YOLO model improves the mAP0.5 by 9.3% over YOLOv8n, and its performance on the SSDD and RSOD datasets also confirms its superior generalization capabilities. These results confirm the effectiveness and significant progress of our novel approach to small target detection.

ExplainLFS: Explaining neural architectures for similarity learning from local perturbations in the latent feature space

Despite the increasing development in recent years of explainability techniques for deep neural networks, only some are dedicated to explaining the decisions made by neural networks for similarity learning. While existing approaches can explain classification models, their adaptation to generate visual similarity explanations is not trivial. Neural architectures devoted to this task learn an embedding that maps similar examples to nearby vectors and non-similar examples to distant vectors in the feature space. In this paper, we propose a post-hoc agnostic technique that explains the inference of such architectures on a pair of images. The proposed method establishes a relation between the most important features of the abstract feature space and the input feature space (pixels) of an image. For this purpose, we employ a relevance assignment and a perturbation process based on the most influential latent features in the inference. Then, a reconstruction process of the images of the pair is carried out from the perturbed embedding vectors. This process relates the latent features to the original input features. The results indicate that our method produces “continuous” and “selective” explanations. A sharp drop in the value of the function (summarized by a low value of the area under the curve) indicates its superiority over other explainability approaches when identifying features relevant to similarity learning. In addition, we demonstrate that our technique is agnostic to the specific type of similarity model, e.g., we show its applicability in two similarity learning tasks: face recognition and image retrieval.

OENet: An overexposure correction network fused with residual block and transformer

With the wide application of vision in the fields of autonomous driving and medical imaging, the demand for overexposure image correction algorithms is becoming increasingly urgent. However, existing overexposure image correction algorithms can lead to problems such as blurring, color bias, and over-enhancement of the generated images. Optimizing overexposure image quality has a significant impact on improving system performance, accuracy, and safety. In this paper, we propose an overexposure image correction network. First, we built the Detail Enhancement Module (DEM). It adopts global average pooling on each channel of the input feature map. After pooling, an activation function is used for nonlinear mapping to generate a channel attention weight vector. And it is multiplied with the original input feature map to achieve the purpose of enhancing the details of the overexposed image. Second, we construct a context-aware backbone (CAB) to extract features such as color and texture. The linear attention gating mechanism replaces the multi-head attention module in Transformer, and reduces the computational complexity in high-resolution images while maintaining performance by learning linear transformation and attention gating. Finally, we design an attention-guided feature fusion (AGFF) to fuse shallow and deep features. It computes weight vectors for shallow features through an attention module. The calculated result is converted to the same dimension as the input feature by bilinear interpolation, so as to enrich the semantic information and detailed information of the generated image. In addition to designing the network structure, we design a hybrid loss function to improve the quality of the generated image from the spatial and structural aspects, and the exposure function can correct the exposure degree of the generated image. Experiments are conducted on two public datasets and the dataset in this paper. Specifically, the PSNR and SSIM of images generated on the dataset MSEC increased by 1.3813% and 5.56%. The PSNR and SSIM of images generated on the dataset SICE increased by 1.545% and 4.64%. The proposed method can effectively generate clear and high-fidelity images.

Hierarchical Locality-Aware Deep Dictionary Learning for Classification

Deep dictionary learning (DDL) shows good performance in visual classification tasks. However, almost all existing DDL methods ignore the locality relationships between the input data representations and the learned dictionary atoms, and learn sub-optimal representations in the feature coding stage, which are less conducive to classification. To this end, we propose a hierarchical locality-aware deep dictionary learning (HILADLE) framework for classification, which can learn locality-constrained dictionaries at different abstract levels through hierarchical dictionary learning. The locality constraints play an important role in learning informative dictionary atoms while preserving the data structure in the original input feature space. Moreover, instead of using an identity activation function like existing DDL methods, we further boost the generalization performance of our HILADLE method with a ReLU activation function to deal with the overfitting issue caused by over-parameterization, inspired by its effectiveness in deep neural networks. Finally, the concatenation of all feature representations learned at different layers is used as input to the final classifier. We demonstrate, through an extensive set of experiments on several benchmark face recognition, image classification, and age estimation datasets, that our method is able to surpass several dictionary learning, deep dictionary learning and deep learning methods.

A Multi-Target Regression Method to Predict Element Concentrations in Tomato Leaves Using Hyperspectral Imaging.

Recent years have seen the development of novel, rapid, and inexpensive techniques for collecting plant data to monitor the nutritional status of crops. These techniques include hyperspectral imaging, which has been widely used in combination with machine learning models to predict element concentrations in plants. When there are multiple elements, the machine learning models are trained with spectral features to predict individual element concentrations; this type of single-target prediction is known as single-target regression. Although this method can achieve reliable accuracy for some elements, there are others that remain less accurate. We aimed to improve the accuracy of element concentration predictions by using a multi-target regression method that sequentially augmented the original input features (hyperspectral imaging) by chaining the predicted element concentration values. To evaluate the multi-target method, the concentrations of 17 elements in tomato leaves were predicted and compared with the single-target regression results. We trained 5 machine learning models with hyperspectral data and predicted element concentration values and found a significant improvement in the prediction accuracy for 10 elements (Mg, P, S, Mn, Fe, Co, Cu, Sr, Mo, and Cd). Furthermore, our multi-target regression method outperformed single-target predictions by increasing the coefficient of determination (R2) for elements such as Mn, Cu, Co, Fe, and Mg by 12.5%, 10.3%, 11%, 10%, and 8.4%, respectively. Hence, our multi-target method can improve the accuracy of predicting 10-element concentrations compared to single-target regression.

Prediction of plant LncRNA-protein interactions based on feature fusion and an improved residual network

LncRNA(long non-coding RNA)-protein interaction (LPI) has effects on chromosome structure and gene transcription, participating in key cellular processes such as signal transduction, chromosome replication, material transport, and mitosis. Accurate identification of LPIs will provide scientific basis for understanding the molecular mechanisms of LncRNA-related diseases, thereby promoting the progress of disease diagnosis technologies and the development of therapeutic procedures to a certain extent. How to comprehensively mine feature information reflecting functional attributes from LncRNAs and proteins, apply deep learning to extract advanced features from the original input features of proteins and LncRNAs, and effectively fuse them have been important research challenges in this field. Aiming at the limitations of existing methods, based on deep learning techniques such as bi-directional long short-term memory (BiLSTM), attention mechanism, and an improved residual network, an LPI prediction model called LPI-LSTM-ResNet is constructed in this paper. Firstly, the sequence and structural information of LncRNAs and proteins are extracted from different perspectives. Then, the deep interaction fused features between LncRNAs and proteins are obtained by BiLSTM and attention mechanism. Finally, the fused features are input into an improved residual network with LSTM as the residual element to preserve long-distance dependencies between sequences. The 5-fold cross-validation results indicate that the feature combination strategy, feature fusion strategy, and the improved residual network consistently improve the LPI prediction performance. Compared with existing methods on the same plant datasets, LPI-LSTM-ResNet exhibits superior performance.

Lightweight 3D hand pose estimation by cascading CNNs with reinforcement learning

This paper proposes a novel strategy for lightweight 3D hand pose estimation. The strategy decomposes the estimation process into feature extraction and feature exploitation, where feature extraction performs dimension reduction on the original input and outputs feature vectors. Feature exploitation is further analyzed and considered as a path optimization problem, and reinforcement learning (RL) is proved to be capable of tackling the problem accurately. A framework cascading convolutional neural networks (CNNs) and RL is next introduced to validate the effectiveness of the proposed strategy, where two different backbones are used to extract features, and RL is extended into continuous space to enhance accuracy. Ablation studies and experiments are carried out on NYU and ICVL datasets using the proposed strategy with continuous RL. The results show that the accuracy of continuous RL exceeds discrete RL, and the rapidity and accuracy leads the backbones. Comparative studies show the strategy achieves leading rapidity and accuracy in single-view depth-based methods.

Deep feature screening: Feature selection for ultra high-dimensional data via deep neural networks

The applications of traditional statistical feature selection methods to high-dimension, low-sample-size data often struggle and encounter challenging problems, such as overfitting, curse of dimensionality, computational infeasibility, and strong model assumptions. In this paper, we propose a novel two-step nonparametric approach called Deep Feature Screening (DeepFS) that can overcome these problems and identify significant features with high precision for ultra high-dimensional, low-sample-size data. This approach first extracts a low-dimensional representation of input data and then applies feature screening on the original input feature space based on multivariate rank distance correlation recently developed by Deb and Sen (2021). This approach combines the strengths of both deep neural networks and feature screening, thereby having the following appealing features in addition to its ability of handling ultra high-dimensional data with small number of samples: (1) it is model free and distribution free; (2) it can be used for both supervised and unsupervised feature selection; and (3) it is capable of recovering the original input data. The superiority of DeepFS is demonstrated via extensive simulation studies and real data analyses.

Two-stage label distribution learning with label-independent prediction based on label-specific features

Label distribution learning (LDL) explicitly models label ambiguity based on the degree to which each label describes an instance. Existing LDL algorithms typically consider all features shared by all class labels. However, irrelevant or redundant information in the feature space can degrade the performance of the model. In this study, a novel two-stage LDL algorithm with label-independent prediction based on label-specific features (TSLIP) is proposed. The first stage combines sparse feature selection and instance correlation constraints. The original high-dimensional input feature space is mapped to a derived low-dimensional label-independent prediction space. The mapping operation not only supports the extraction of label-specific features but also brings the derived space closer to the label distribution space. The second stage employs the Pearson coefficient to explore the label correlation to enhance the performance of the model. Subsequently, the label-independent predictions are normalized to obtain the final label distribution. Extensive experiments are conducted on twelve real-world datasets. The results of the seven measures demonstrate the superiority of our algorithm compared with several state-of-the-art methods.

EGA-Net: Edge feature enhancement and global information attention network for RGB-D salient object detection

With the supplement of texture and geometry cues in depth maps, salient object detection (SOD) shifts from 2D to 3D, aiming to detect the most attractive object in a pair of color and depth images. Previous work primarily focused on regional integrity. Few methods are used to improve the edge quality of prediction results, resulting in a final prediction with a complete structure but blurred edges. Moreover, due to the complexity of real-life scenarios, the problem of effectively separating the salient object from complex background has become a hot potato. Aiming to address these issues, we propose a novel network, EGA-Net, to improve the edge quality and highlight the main features of the salient object. Specifically, in the EGA-Net, we propose a feature interaction (FI) module and an edge feature enhancement (EFE) module, respectively. Among them, the FI module is used to remove unimodal feature redundancy, capture multi-modal feature complementarity, and reduce the contamination of low-quality depth maps. The EFE is used to improve the edge quality of the final salient object prediction results. Furthermore, a Global Information Guide Integration (GIGI) module has been proposed to suppress the background noise and effectively highlight the salient objects’ main features. It uses interleaving and fusion methods to automatically select and enhance the vital information in the original input features under the guidance of global features. We put the training of EGA-Net under the supervision of a new hybrid loss function that can simultaneously take global pixel point, foreground, and depth map quality into account. Quantitative and qualitative experiment results demonstrate that our method outperforms the 19 advanced methods on eight publicly available RGB-D salient object detection datasets with five evaluation metrics. You can find the code and results of our method athttps://github.com/guanyuzong/EGA-Net.

Prediction model of sparse autoencoder-based bidirectional LSTM for wastewater flow rate.

Sanitary sewer overflows caused by excessive rainfall derived infiltration and inflow is the major challenge currently faced by municipal administrations, and therefore, the ability to correctly predict the wastewater state of the sanitary sewage system in advance is especially significant. In this paper, we present the design of the Sparse Autoencoder-based Bidirectional long short-term memory (SAE-BLSTM) network model, a model built on Sparse Autoencoder (SAE) and Bidirectional long short-term memory (BLSTM) networks to predict the wastewater flow rate in a sanitary sewer system. This network model consists of a data preprocessing segment, the SAE network segment, and the BLSTM network segment. The SAE is capable of performing data dimensionality reduction on high-dimensional original input feature data from which it can extract sparse potential features from the aforementioned high-dimensional original input feature data. The potential features extracted by the SAE hidden layer are concatenated with the smooth historical wastewater flow rate features to create an augmented previous feature vector that more accurately predicts the wastewater flow rate. These augmented previous features are applied to the BLSTM network to predict the future wastewater flow rate. Thus, this network model combines two kinds of abilities, SAE's low-dimensional nonlinear representation for original input feature data and BLSTM's time series prediction for wastewater flow rate. Then, we conducted extensive experiments on the SAE-BLSTM network model utilizing the real-world hydrological time series datasets and employing advanced SVM, FCN, GRU, LSTM, and BLSTM models as comparison algorithms. The experimental results show that our proposed SAE-BLSTM model consistently outperforms the advanced comparison models. Specifically, we selected a 3 months period training dataset in our dataset to train and test the SAE-BLSTM network model. The SAE-BLSTM network model yielded the lowest RMSE, MAE, and highest R2, which are 242.55, 179.05, and 0.99626, respectively.

Direct domain adaptation through reciprocal linear transformations.

We propose a direct domain adaptation (DDA) approach to enrich the training of supervised neural networks on synthetic data by features from real-world data. The process involves a series of linear operations on the input features to the NN model, whether they are from the source or target distributions, as follows: (1) A cross-correlation of the input data (i.e., images) with a randomly picked sample pixel (or pixels) of all images from the input or the mean of all randomly picked sample pixel (or pixels) of all input images. (2) The convolution of the resulting data with the mean of the autocorrelated input images from the other domain. In the training stage, as expected, the input images are from the source distribution, and the mean of auto-correlated images are evaluated from the target distribution. In the inference/application stage, the input images are from the target distribution, and the mean of auto-correlated images are evaluated from the source distribution. The proposed method only manipulates the data from the source and target domains and does not explicitly interfere with the training workflow and network architecture. An application that includes training a convolutional neural network on the MNIST dataset and testing the network on the MNIST-M dataset achieves a 70% accuracy on the test data. A principal component analysis (PCA), as well as t-SNE, shows that the input features from the source and target domains, after the proposed direct transformations, share similar properties along the principal components as compared to the original MNIST and MNIST-M input features.

An intelligent and multi-channel deep feature enhanced framework for predicting difficult-to-measure CTQ

Some CTQs (Critical-To-Quality) are difficult to measure online due to economic or technical limitations, hindering the timely decision-making to promote the production operation. Meanwhile, existing CTQ prediction methods may have inferior performance when considering the specific characteristic of data under the new industrial revolution context. Fortunately, various operating data have been increasingly collected by widely deployed sensors, enabling the online measurement of numerous operating parameters and partial CTQs. In addition, deep exploration of these data can be achieved with the assistance of emerging data analytic techniques. Hence, an integrated framework is proposed to predict difficult-to-measure CTQs using computational intelligence models, in which the multi-channel deep features could be extracted sequentially to enhance the prediction performance. Specifically, a feature extractor is first constructed to obtain deep time-dependent features using Bidirectional Gated Recurrent Unit. These extracted features describe the dependency between the easy-to-measure CTQ and operating parameters. Then, these features and original input features are concatenated to enrich the available information for predicting difficult-to-measure CTQs. To reduce the risk of high-dimensionality and remove redundant information introduced by the concatenated operation, Kernel Principal Component Analysis is further employed to obtain abstract features. Next, an inverse label propagation strategy is proposed to find matched feature-CTQ pairs to alleviate the unmatched problem caused by data imbalance. Afterwards, Random Forest, which has the dual advantages of ensemble learning and intelligence algorithm, is used to build the feature-CTQ function and achieve CTQ prediction. Finally, the effectiveness and superiority of the proposed model are verified by forecasting tensile strength characteristic for a commutator welding process.

A TMA-Seq2seq Network for Multi-Factor Time Series Sea Surface Temperature Prediction

Sea surface temperature (SST) is closely related to global climate change, ocean ecosystem, and ocean disaster. Accurate prediction of SST is an urgent and challenging task. With a vast amount of ocean monitoring data are continually collected, data-driven methods for SST time-series prediction show promising results. However, they are limited by neglecting complex interactions between SST and other ocean environmental factors, such as air temperature and wind speed. This paper uses multi-factor time series SST data to propose a sequence-to-sequence network with two-module attention (TMA-Seq2seq) for long-term time series SST prediction. Specifically, TMA-Seq2seq is an LSTM-based encoder-decoder architecture facilitated by factor- and temporal-attention modules and the input of multi-factor time series. It takes six-factor time series as the input, namely air temperature, air pressure, wind speed, wind direction, SST, and SST anomaly (SSTA). A factor attention module is first designed to adaptively learn the effect of different factors on SST, followed by an encoder to extract factor-attention weighted features as feature representations. And then, a temporal attention module is designed to adaptively select the hidden states of the encoder across all time steps to learn more robust temporal relationships. The decoder follows the temporal-attention module to decode the feature vector concatenated from the weighted features and original input feature. Finally, we use a fully-connect layer to map the feature into prediction results. With the two attention modules, our model effectively improves the prediction accuracy of SST since it can not only extract relevant factor features but also boost the long-term dependency. Extensive experiments on the datasets of China Coastal Sites (CCS) demonstrate that our proposed model outperforms other methods, reaching 98.29% in prediction accuracy (PACC) and 0.34 in root mean square error (RMSE). Moreover, SST prediction experiments in China’s East, South, and Yellow Sea site data show that the proposed model has strong robustness and multi-site applicability.

Mining of Weak Fault Information Adaptively Based on DNN Inversion Estimation for Fault Diagnosis of Rotating Machinery

Fault feature extraction plays a significant role in bearing fault diagnosis, especially in incipient fault period. Recently, deep neural network has been favored by many researchers due to its excellent hierarchical feature extraction capabilities. The existing diagnosis methods based on deep neural network mostly take original condition monitoring data as input and further convert fault diagnosis into pattern recognition issues. Although it improves the level of intelligent diagnosis, it is still confronted with practical problems. Since deep neural network includes many non-linear mapping layers, the extracted fault-related features show a high degree of abstraction, which reduces the practical level of this technology. Aimed at this problem, a mining method of understandably weak fault information is proposed based on deep neural network inversion estimation. From perspective of neuron response degree, this method mines the most sensitive input pattern that maximizes the neuron’s activation value of network output layer in the original input feature space. It can intuitively mine the weak fault information from the original signal. Two bearing experiments verify the effectiveness and reliability of the proposed method.

Broad Autoencoder Features Learning for Classification Problem

Activation functions such as Tanh and Sigmoid functions are widely used in Deep Neural Networks (DNNs) and pattern classification problems. To take advantages of different activation functions, the Broad Autoencoder Features (BAF) is proposed in this work. The BAF consists of four parallel-connected Stacked Autoencoders (SAEs) and each of them uses a different activation function, including Sigmoid, Tanh, ReLU, and Softplus. The final learned features can merge such features by various nonlinear mappings from original input features with such a broad setting. This helps to excavate more information from the original input features. Experimental results show that the BAF yields better-learned features and classification performances.

Sparse random feature maps for the item-multiset kernel

Random feature maps are a promising tool for large-scale kernel methods. Since most random feature maps generate dense random features causing memory explosion, it is hard to apply them to very-large-scale sparse datasets. The factorization machines and related models, which use feature combinations efficiently, scale well for large-scale sparse datasets and have been used in many applications. However, their optimization problems are typically non-convex. Therefore, although they are optimized by using gradient-based iterative methods, such methods cannot find global optimum solutions in general and require a large number of iterations for convergence. In this paper, we define the item-multiset kernel, which is a generalization of the itemset kernel and dot product kernels. Unfortunately, random feature maps for the itemset kernel and dot product kernels cannot approximate the item-multiset kernel. We thus develop a method that converts an item-multiset kernel into an itemset kernel, enabling the item-multiset kernel to be approximated by using a random feature map for the itemset kernel. We propose two random feature maps for the itemset kernel, which run faster and are more memory efficient than the existing feature map for the itemset kernel. They also generate sparse random features when the original (input) feature vector is sparse and thus linear models using proposed methods . Experiments using real-world datasets demonstrated the effectiveness of the proposed methodology: linear models using the proposed random feature maps ran from 10 to 100 times faster than ones based on existing methods.

Broad Autoencoder Features Learning for Classification Problem

Activation functions such as Tanh and Sigmoid functions are widely used in Deep Neural Networks (DNNs) and pattern classification problems. To take advantages of different activation functions, the Broad Autoencoder Features (BAF) is proposed in this work. The BAF consists of four parallel-connected Stacked Autoencoders (SAEs) and each of them uses a different activation function, including Sigmoid, Tanh, ReLU, and Softplus. The final learned features can merge such features by various nonlinear mappings from original input features with such a broad setting. This helps to excavate more information from the original input features. Experimental results show that the BAF yields better-learned features and classification performances.